Blog  /   ยท  8 min read

MCP in plain English: what it is, where it helps, and how to set it up in ServiceNow

One protocol lets AI assistants use your tools and data without a custom integration for every pair. Here is how it works and how to turn it on in your instance.

MCPServiceNowAgentic AI

What MCP is

The Model Context Protocol (MCP) is an open standard that defines how an AI application talks to external tools and data. Before it, every assistant needed its own connector for every system. With MCP, a system exposes its capabilities once, and any MCP-aware assistant can use them.

Think of it as USB-C for AI tools. Two roles matter:

  • MCP client: the AI app, such as a chat assistant or an agent framework. It decides when to call a tool.
  • MCP server: the system that publishes tools, like "get case summary" or "search knowledge". It runs them and returns results.
AI clientasksMCP serverrunsYour system

A session is short. The client connects and authenticates, asks the server which tools exist, then calls the ones it needs with structured inputs.

// 1. Discover
{ "jsonrpc": "2.0", "id": 1, "method": "tools/list" }

// 2. Call a tool
{ "jsonrpc": "2.0", "id": 2, "method": "tools/call",
  "params": { "name": "get_case_summary",
              "arguments": { "case_number": "CS0012345" } } }

Each tool carries a name, a plain-language description and an input schema. The model reads those to decide what to call, so clear descriptions matter as much as the code behind them.

Where it helps

MCP pays off wherever an AI needs live, governed access to a business system instead of a pasted snippet.

  • Customer support: summarize a case, find related knowledge articles, draft a reply from case history.
  • IT operations: look up an incident, check recent changes on a service, suggest next steps for triage.
  • HR and employee service: answer policy questions from approved knowledge and open a request on someone's behalf.
  • Developer assistants: let a coding tool query records or call an internal API while you work.
  • Multi-system workflows: one agent that reads from ServiceNow, then acts in another tool, using the same protocol for both.

The win is reuse. A tool you expose once is available to every compatible client, and access control stays in one place.

Set it up in ServiceNow

ServiceNow ships this as the MCP Server Console. You create a server, add tools to it, and let clients connect over OAuth 2.0. Menu names can differ slightly between releases, so check the documentation for yours.

Before you start

  • The Now Assist application is activated, with the Generative AI Controller and the Now Assist plugins installed. MCP Server Console is available through the ServiceNow Store and depends on them.
  • You have the AI administrator role on the instance.
  • You know which capability you want to expose and who may use it.
  1. Open MCP Server Console and create a server

    Give it a clear name and description. Use one server per audience or workflow, for example one for IT and one for HR, so each client only sees what it needs.

  2. Add tools

    Tools can be built from Now Assist skills, subflows, actions, REST APIs, scripted APIs and knowledge graphs. Pick the capability, then choose which fields appear as tool inputs. Expose the minimum.

  3. Set up authentication

    Create an OAuth inbound integration for each client. Newer releases embed OAuth client registration in the server wizard. Third-party identity providers are also supported in recent releases.

  4. Connect the client

    In your MCP client, add the server URL and the OAuth details. The URL usually follows this shape, but copy the exact value from the console.

    https://<your-instance>.service-now.com/sncapps/mcp-server/mcp/<server_name>
  5. Test discovery, then a real call

    Ask the client to list tools. If you see your tools, call one with sample input and check the result matches what the same user would see in the platform.

When tools do not show up

The most common symptom is a client that authenticates but sees an empty tool list. Check these in order:

  • Wrong server URL. The client points at a different server than the one you added tools to.
  • Tool not attached or not active. Open the tool in the console and confirm it is active and belongs to this server.
  • OAuth scope or user permissions. A valid token can still return no tools if the integration or the user lacks access.

Design tips

  • Write descriptions for the model. Say what the tool does, what it needs and what it returns. Vague names lead to wrong calls.
  • Keep tools narrow. "Get case summary" beats "do case stuff". Small tools are easier to secure and test.
  • Apply least privilege. Separate servers and scoped OAuth clients per use case. Prefer read tools first, then add write actions deliberately.
  • Return compact, structured results. Models handle trimmed fields better than full records.
  • Log and review calls. Treat tool use like any other API traffic you need to audit.

MCP does not make an AI smarter. It gives it safe, standard reach into the systems where the work already happens. That is where most of the value is.

Back to all posts